Limit frame size in pixels to about 16MP, while allowing the fuzzer to test
frame widths and heights above 4096.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Calling dav1d_get_picture() again after it has already returned with
an error due to a memory allocation failure could result in crashes.

Although doing so is not a proper API usage, and the outcome is going
to be unpredictable, we should at least try to avoid crashing.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

dav1d_submit_frame() could erroneously return 0 when tile data memory
allocation failed.

Fixes an assertion failure in dav1d_parse_obus().

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

In the (very unlikely) scenario of a pthread mutex/cond init failure
in the tile state reallocation code some newly allocated mutexes/conds
could leak.

B Krishnan Iyer authored 5 years ago and Martin Storsjö committed 5 years ago

	                     A73             A53
	                Earlier	Now	Earlier	Now

blend_v_w2_8bpc_neon:	122.1	121.3	195.5	195.5
blend_v_w4_8bpc_neon:	248.2	247.5	375.6	358.5
blend_v_w8_8bpc_neon:	210.3	205.2	375.6	358.5
blend_v_w16_8bpc_neon:	252.7	237.1	579.2	590.5
blend_v_w32_8bpc_neon:	347	345.8	997.4	994.1

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Avoid allocating significantly more memory than what is actually used.

Henrik Gramner authored 5 years ago and Henrik Gramner committed 5 years ago

Also eliminate some pointer chasing by allocating tile context buffers
as part of the struct instead of having the struct contain pointers to
separately allocated buffers.

'build_' prefix is reserved by meson, this will become an error in the
future, as indicated by a warning when configuring the build dir.

Closes #285.

For the cdef_filter tests, one could also extend the buffer to
contain 16*11 pixels, to simplify printing it as one rectangular
section.

Extend the common hex_dump function to allow dumping to an arbitrary
FILE* pointer, to reuse it for printing the source pixel buffer in
case of errors.

Martin Storsjö authored 5 years ago and Jean-Baptiste Kempf committed 5 years ago

The speedup for most non-dc-only dct functions is around 9-12x
over the C code generated by GCC 7.3.

Relative speedups vs C for a few functions:

                                              Cortex A53    A72    A73
inv_txfm_add_4x4_dct_dct_0_8bpc_neon:               3.90   4.16   5.65
inv_txfm_add_4x4_dct_dct_1_8bpc_neon:               7.20   8.05  11.19
inv_txfm_add_8x8_dct_dct_0_8bpc_neon:               5.09   6.73   6.45
inv_txfm_add_8x8_dct_dct_1_8bpc_neon:              12.18  10.80  13.05
inv_txfm_add_16x16_dct_dct_0_8bpc_neon:             7.31   9.35  11.17
inv_txfm_add_16x16_dct_dct_1_8bpc_neon:            14.36  13.06  15.93
inv_txfm_add_16x16_dct_dct_2_8bpc_neon:            11.00  10.09  12.05
inv_txfm_add_32x32_dct_dct_0_8bpc_neon:             4.41   5.40   5.77
inv_txfm_add_32x32_dct_dct_1_8bpc_neon:            13.84  13.81  18.04
inv_txfm_add_32x32_dct_dct_2_8bpc_neon:            11.75  11.87  15.22
inv_txfm_add_32x32_dct_dct_3_8bpc_neon:            10.20  10.40  13.13
inv_txfm_add_32x32_dct_dct_4_8bpc_neon:             9.01   9.21  11.56
inv_txfm_add_64x64_dct_dct_0_8bpc_neon:             3.84   4.82   5.28
inv_txfm_add_64x64_dct_dct_1_8bpc_neon:            14.40  12.69  16.71
inv_txfm_add_64x64_dct_dct_4_8bpc_neon:            10.91   9.63  12.67

Some of the specialcased identity_identity transforms for 32x32
give insane speedups over the generic C code:

inv_txfm_add_32x32_identity_identity_0_8bpc_neon: 225.26 238.11 247.07
inv_txfm_add_32x32_identity_identity_1_8bpc_neon: 225.33 238.53 247.69
inv_txfm_add_32x32_identity_identity_2_8bpc_neon:  59.60  61.94  64.63
inv_txfm_add_32x32_identity_identity_3_8bpc_neon:  26.98  27.99  29.21
inv_txfm_add_32x32_identity_identity_4_8bpc_neon:  15.08  15.93  16.56

This keeps the put/prep functions close to the 8tap/bilin functions
that use them.

clock_gettime() is only available since MacOS X 10.12 (Sierra).

B Krishnan Iyer authored 6 years ago and B Krishnan Iyer committed 5 years ago

	                A73	A53

blend_h_w2_8bpc_c:	149.3	246.8
blend_h_w2_8bpc_neon:	74.6	137
blend_h_w4_8bpc_c:	251.6	409.8
blend_h_w4_8bpc_neon:	66	146.6
blend_h_w8_8bpc_c:	446.6	844.1
blend_h_w8_8bpc_neon:	68.6	131.2
blend_h_w16_8bpc_c:	830	1513
blend_h_w16_8bpc_neon:	85.9	192
blend_h_w32_8bpc_c:	1605.2	2847.8
blend_h_w32_8bpc_neon:	149.8	357.6
blend_h_w64_8bpc_c:	3304.8	5515.5
blend_h_w64_8bpc_neon:	262.8	629.5
blend_h_w128_8bpc_c:	7895.1	13260.6
blend_h_w128_8bpc_neon:	577	1402
blend_v_w2_8bpc_c:	241.2	410.8
blend_v_w2_8bpc_neon:	122.1	196.8
blend_v_w4_8bpc_c:	874.4	1418.2
blend_v_w4_8bpc_neon:	248.5	375.9
blend_v_w8_8bpc_c:	1550.5	2514.7
blend_v_w8_8bpc_neon:	210.8	376
blend_v_w16_8bpc_c:	2925.3	5086
blend_v_w16_8bpc_neon:	253.4	608.3
blend_v_w32_8bpc_c:	5686.7	9470.5
blend_v_w32_8bpc_neon:	348.2	994.8
blend_w4_8bpc_c:	201.5	309.3
blend_w4_8bpc_neon:	38.6	99.2
blend_w8_8bpc_c:	531.3	944.8
blend_w8_8bpc_neon:	55.1	125.8
blend_w16_8bpc_c:	1992.8	3349.8
blend_w16_8bpc_neon:	150.1	344
blend_w32_8bpc_c:	4982	8165.9
blend_w32_8bpc_neon:	360.4	910.9

Luca Barbato authored 5 years ago and Luca Barbato committed 5 years ago

Limited to PowerPC64 LE for now.

On older versions of glibc, clock_gettime isn't available in the main
libc, but part of a separate librt.

Only look for librt if clock_gettime isn't available otherwise.

Martin Storsjö authored 5 years ago and Janne Grunau committed 5 years ago

This fixes building with raspbian compilers, that default to armv6.
The isb instruction is unavailable on armv6, and the cycle counter
register is accessed differently there as well.

This fixes issue #282.

nasm -v can actually fail for example on macOS, where nasm could be a
stub executable that forwards commands to the real nasm, but if the real
nasm is not installed, fails.
This would lead to a confusing error message due to the out of bounds
array access, to avoid that, explicitly check the exit code.